Xiaomi

Bonsai 27B

Bonsai 27B is an open-weights 27B-parameter model by Xiaomi. It holds an Intelligence score of 59 on the local.ai leaderboard (2026-08 snapshot).

open weights
59
local.ai Intelligence
Parameters
27B
Benchmarks (local.ai snapshot)
τ²-bench
68%
GAIA
59%
GDPval
58%
HuggingFace metadata · updated 2026-07-17
repo
prism-ml/Bonsai-27B-gguf
license
apache-2.0
downloads
1,801,896
likes
776

Metadata refreshed from the HuggingFace API (snapshot 2026-08-15T15:53:42.405Z). Context length and long-form descriptions are the next pipeline stage.

Does it fit your hardware?

Detect or select your GPU first. Green = weights + context headroom, yellow = tight, grey = does not fit.

Q4_K_M 15.7GQ6_K 24.3GQ8_0 29.7GFP16 54G
Quantization≈ File sizeNote
IQ2_XXS7.3 GBExtreme compression, largest quality loss
IQ3_XXS9.7 GBVery small
Q3_K_M12.4 GBSmall
NVFP413.5 GBNVIDIA 4-bit
Q4_K_S14.6 GBCommon
Q4_K_M15.7 GBMost common sweet spot
Q5_K_M20.5 GBBalanced
Q6_K24.3 GBNear-lossless
Q8_029.7 GBNear-lossless
FP1654 GBOriginal weights

Actual files on HuggingFace

FileSize
Bonsai-27B-F16.gguf50.11 GB
Bonsai-27B-dspark-bf16.gguf6.79 GB
Bonsai-27B-Q1_0.gguf3.54 GB
Bonsai-27B-dspark-Q4_1.gguf1.66 GB
Bonsai-27B-mmproj-BF16.gguf0.87 GB
Bonsai-27B-mmproj-Q8_0.gguf0.59 GB
GGUF: prism-ml/Bonsai-27B-gguf · 1,801,896 downloads
FileSize
Bonsai-27B-F16.gguf50.11 GBdownload
Bonsai-27B-Q1_0.gguf3.54 GBdownload
Bonsai-27B-dspark-Q4_1.gguf1.66 GBdownload
Bonsai-27B-dspark-bf16.gguf6.79 GBdownload
Bonsai-27B-mmproj-BF16.gguf0.87 GBdownload
Bonsai-27B-mmproj-Q8_0.gguf0.59 GBdownload
GGUF: prism-ml/Ternary-Bonsai-27B-gguf · 775,667 downloads
FileSize
Ternary-Bonsai-27B-F16.gguf50.11 GBdownload
Ternary-Bonsai-27B-PQ2_0.gguf6.67 GBdownload
Ternary-Bonsai-27B-Q2_0.gguf6.67 GBdownload
Ternary-Bonsai-27B-Q2_g64.gguf7.06 GBdownload
Ternary-Bonsai-27B-dspark-Q4_1.gguf1.81 GBdownload
Ternary-Bonsai-27B-dspark-bf16.gguf6.79 GBdownload
Ternary-Bonsai-27B-mmproj-BF16.gguf0.87 GBdownload
Ternary-Bonsai-27B-mmproj-Q8_0.gguf0.59 GBdownload

What you save self-hosting this model

Uses the model's own API price when it exists; otherwise the closest commercial equivalent.

Waiting for hardware detection…

RunLocal presents third-party leaderboard data (local.ai / Exo Labs, 2026-08) with its own fit and cost estimates. Verify pricing, licenses and file sizes at the source before deploying a model.